Table of contents

1. Initial Discussion - Josiah
2. Motion Control Work - Josiah & Adam
   1. Mujoco Implementation
   2. Walk Engine Integration
3. Self-Play Implementation - Josiah & Will
4. Multi-Agent Implementation - Yuhao
5. AbstractSim V2 Status - Josiah & Abhinav
6. Implementation Progress - Josiah & Edison
7. Vision Recognition - Zisen & Josiah
   1. Data Collection
   2. Implementation Priority
8. Goalkeeper Development - Josiah & Parv
9. Meeting Wrap-up - Josiah
   1. Timeline
   2. Spring Semester Focus

Initial Discussion - Josiah

• Discussed merging diverged simulator versions
• Need to check status with team members
• Goal: merge versions to establish main v2 simulator

Motion Control Work - Josiah & Adam

Mujoco Implementation

• Adam working on Mujoco motion control
• Options for BHuman motion control:
  • Behavior cloning approach
  • Direct code implementation
• Considering Isaac Gym with PPO/SAC training

Walk Engine Integration

• Discussed porting BHuman walk to Mujoco
• Proposal: implement as standalone binary with velocity commands
• Challenges:
  • Walk cloning complexity
  • Managing internal state variables
• Possible solution: recurrent neural network approach
• Critical timing: Changes needed by February/March

Self-Play Implementation - Josiah & Will

• Will exploring RLlib for self-play
• Implementation challenges:
  • Example not working out of box
  • Performance issues with multiple policies
• Discussing stable baselines integration

Multi-Agent Implementation - Yuhao

• Provided insights on multi-agent approaches
• Suggested reviewing Tianshou’s implementation

AbstractSim V2 Status - Josiah & Abhinav

• Fixed push ball to goal policy issues
• Need to verify transfer performance
• Working on main V2 branch development

Implementation Progress - Josiah & Edison

• Edison working with latest commit
• Current challenges:
  • Dimension mismatch errors
  • Policy integration issues

Vision Recognition - Zisen & Josiah

Data Collection

• ETH team cannot share data (privacy concerns)
• Exploring alternatives:
  • Unreal Engine
  • SimRobot for data collection

Implementation Priority

• Focus on NAO robot deployment
• Direct hardware implementation preferred over simulation
• Network architecture appears suitable based on literature

Goalkeeper Development - Josiah & Parv

• Parv working on goalkeeper implementation
• Recommendations:
  • Use existing stable baselines
  • Avoid new RL algorithm implementation
• Arranged mentoring with Abhinav for V2 setup

Meeting Wrap-up - Josiah

Timeline

• Approaching semester end
• RoboCup application due after winter break

Spring Semester Focus

• Emphasis on practical implementations
• Key areas:
  • Multi-agent aspects
  • V2 migration
  • Practical applicability of research